Value Augmented Sampling for Language Model Alignment and Personalization

TL;DR

This paper introduces Value Augmented Sampling (VAS), a novel reward optimization framework that enhances LLM alignment and personalization by maximizing reward functions without modifying the model weights, reducing inference costs and enabling flexible reward composition.

Contribution

VAS provides a stable, efficient method for reward optimization in LLMs that does not require co-training the value function or access to model weights, facilitating adaptation of API-only models like ChatGPT.

Findings

VAS outperforms PPO and DPO on standard benchmarks.

Achieves results comparable to Best-of-128 with lower inference cost.

Enables reward composition and personalized control during deployment.

Abstract

Aligning Large Language Models (LLMs) to cater to different human preferences, learning new skills, and unlearning harmful behavior is an important problem. Search-based methods, such as Best-of-N or Monte-Carlo Tree Search, are performant, but impractical for LLM adaptation due to their high inference cost. On the other hand, using Reinforcement Learning (RL) for adaptation is computationally efficient, but performs worse due to the optimization challenges in co-training the value function and the policy. We present a new framework for reward optimization, Value Augmented Sampling (VAS), that can maximize different reward functions using data sampled from only the initial, frozen LLM. VAS solves for the optimal reward-maximizing policy without co-training the policy and the value function, making the optimization stable, outperforming established baselines, such as PPO and DPO, on standard benchmarks, and achieving comparable results to Best-of-128 with lower inference cost. Unlike existing RL methods that require changing the weights of the LLM, VAS does not require access to the weights of the pre-trained LLM. Thus, it can even adapt LLMs (e.g., ChatGPT), which are available only as APIs. In addition, our algorithm unlocks the new capability of composing several rewards and controlling the extent of each one during deployment time, paving the road ahead for the future of aligned, personalized LLMs.